The 'Bigfoot Effect' in AI Search: Why 20% Fewer Brands Get Cited by ChatGPT Than Last Month

On March 4, 2026, OpenAI quietly switched ChatGPT's default model from GPT-4o/5.2 to GPT-5.3 Instant. The user experience barely changed. But beneath the surface, something shifted that affects every brand with an online presence: the number of unique domains cited per response dropped from 19 to 15, a 20% decline in just one model update.

A new study from Meteoria, published by Search Engine Land on May 15, tracked this shift with scientific precision. Over 14 weeks, the research team ran 400 daily prompts across multiple model versions, documenting exactly how ChatGPT's citation behavior changed. Their findings have a name: the "Bigfoot Effect," borrowed from a 2012 Google update where a single domain dominated search results. In AI search, the same concentration is happening, but the mechanics are entirely different.

The Data: 20% Fewer Brands, Same Response Length

The core finding is stark. Between mid-January and late April 2026, the average number of unique domains cited per ChatGPT response fell from 19 to 15. Unique URLs per response dropped from 24 to 19. But the URL-to-domain ratio held steady at 1.26, meaning ChatGPT is not citing fewer pages per domain. It is citing fewer domains overall.

This is not ChatGPT giving shorter answers. It is ChatGPT drawing from a smaller pool of sources while maintaining the same depth per source. The pie is the same size. The slices are fewer.

For brands that depend on AI visibility, whether ChatGPT recommends their product, cites their research, or mentions their name in responses, this is the most important data point of 2026. If your brand was source number 18 or 19 in a response before March 4, you may simply no longer appear.

The web.run Reverse Engineering

The Meteoria study goes beyond surface metrics. The team reverse-engineered ChatGPT's internal search system, called web.run, and documented 12 distinct operations the tool performs during a single user query. These include:

• search_query: The initial web search triggered by a user question
• browse_rewritten_queries: Product-specific searches with rewritten queries
• open: Opening a specific URL from search results
• find: Searching within an opened page for relevant content
• click: Following links within an opened page
• screenshot: Capturing visual content from a page
• product_query: Specialized product search with price and availability data

This is not a simple search-and-summarize pipeline. ChatGPT runs 10 or more "fan-out queries" per response in GPT-5.4 Thinking mode, compared to just 2-3 in GPT-5.3. Each fan-out query can target a different source, a different angle, or a different data type. The system is more sophisticated, more thorough, and more selective about which domains it trusts.

Parametric vs. Dynamic Visibility: A New Framework

The study's most actionable contribution is a framework that separates AI visibility into two types:

Parametric visibility is encoded directly in the model's training data. When GPT-5.3 was trained, certain brands, concepts, and relationships were baked into its neural weights. These associations are persistent, model-level, and relatively stable across conversations. The study calls this "E-E-A-T for LLMs": authority that is encoded, not retrieved.

Dynamic visibility comes from real-time web search during a conversation. When ChatGPT uses web.run to fetch current information, it creates a temporary citation surface that depends on the model version, the specific query, and the search results available at that moment. Dynamic visibility is volatile. It can collapse overnight with a model update.

The critical insight: a brand absent from parametric memory will not even be considered as a search candidate. ChatGPT does not randomly discover new brands through web search. It starts with what it already "knows" and uses web search to validate, update, or expand within established boundaries. If your brand is not in the model's parametric memory, web search will not save you.

The site: Operator and Trusted Domain Restriction

One of the most revealing findings in the study involves how GPT-5.4 Thinking uses search operators. The model frequently issues queries with "site:" operators, restricting searches to specific trusted domains. This means ChatGPT is not performing open-ended web searches. It is targeting specific sources it has already identified as authoritative for a given topic.

The implications are significant. If your domain is not in the trusted set for your category, ChatGPT will not even look for you. It will search within its curated list and cite what it finds, which means the same authoritative sources get cited repeatedly while smaller or newer brands are systematically excluded.

Independent log analysis by Jerome Salomon at Oncrawl corroborates this finding. ChatGPT-User, the crawler that fetches actual page content during conversations (separate from OAI-SearchBot, which builds the search index), has settled at a lower crawl volume in recent months. Fewer crawls of a broader set of domains, more concentrated crawling of trusted sources.

Reddit: The Exempt Domain

The study also uncovered a notable asymmetry in ChatGPT's system prompt. Reddit is the only domain explicitly exempted from copyright word limits when ChatGPT fetches and quotes content. Every other source has extraction limits that cap how much content can be pulled. Reddit gets a free pass.

This means Reddit content has disproportionate influence on ChatGPT's responses. Brands that are frequently discussed on Reddit have a parametric visibility advantage that no other platform can match. For brands seeking to improve AI visibility, Reddit presence is not optional. It is structurally advantaged.

The Same Prompt, Different Citations

Perhaps the most unsettling finding for brand owners: the same prompt produces meaningfully different citations across GPT-5.2, GPT-5.3, and GPT-5.4. A brand cited as a top recommendation in one model version may be absent from the response in another. There is no single "AI visibility" state. Your brand is visible or invisible depending on which model version the user is running.

This fragmentation will only increase as OpenAI introduces more model variants tailored to different use cases, price points, and performance profiles. The free plan, used by over 90% of ChatGPT's weekly users, runs a lighter model with fewer search operations and fewer citations. Premium users see a broader citation surface. Your AI visibility is literally different depending on who is asking.

What the Bigfoot Effect Means for Brands

The original Bigfoot update in 2012 saw one domain dominate Google search results, pushing competitors off the first page. The AI version is different in mechanics but similar in outcome: fewer brands share the citation surface in each response. The concentration is driven not by a single dominant source but by the model's narrowing trust set and more selective search behavior.

For brands, the implications fall into three categories:

Brands already in parametric memory: You have a structural advantage. Your priority is maintaining your position through consistent, citable content and active presence on platforms ChatGPT trusts (especially Reddit). Monitor your citation rate across model versions.

Brands with dynamic visibility only: You are vulnerable. A model update can eliminate your visibility overnight. Your priority is building parametric memory through sustained, authoritative content that future model training will ingest. Think in 6-12 month horizons, not 30-day sprints.

Brands with no AI visibility: You are invisible by default. ChatGPT will not discover you through web search unless you first enter its parametric memory. This requires a fundamentally different approach than SEO: you need to become a source that AI models are trained on, not just a page that ranks for keywords.

How to Measure Your Exposure

The first step is understanding where you stand. Check your AI citation rate across major platforms. Look at your server logs for ChatGPT-User and OAI-SearchBot crawl activity. Compare your visibility in ChatGPT versus Perplexity versus Google AI Mode, because each platform has different citation mechanics and different trust sets.

If you have a Trustpilot profile, you already have a citation advantage. A separate study from Seer Interactive analyzing over 800,000 AI responses found that brands with even minimal Trustpilot profiles achieve a 53x higher citation rate than brands without one. Review profiles are one of the fastest ways to improve AI visibility because ChatGPT explicitly draws on review data when evaluating brands.

The Road Ahead

The Bigfoot Effect is not a temporary anomaly. It is a structural feature of how large language models evolve. As models become more sophisticated, they develop stronger priors about which sources to trust. This means the citation surface will continue to concentrate around a smaller set of authoritative domains with each model update.

Brands that invest in parametric visibility now, through authoritative content, review profiles, Reddit presence, and structured data, will compound their advantage with every model update. Brands that rely solely on dynamic visibility will find the ground shrinking beneath them.

The 20% drop happened in a single model switch. The next model update could drop it further. The time to understand your AI visibility is not after the next update. It is now.

---

Want to know if your brand is still visible after the model switch? Run a free AI visibility audit to see which AI engines cite you, how often, and for which queries. Learn more about how AI visibility works and what drives citations in the current landscape.